🚧 Note: this repo is under construction 🚧

Llama-2-Ko 🦙🇰🇷

Llama-2-Ko serves as an advanced iteration of Llama 2, benefiting from an expanded vocabulary and the inclusion of a Korean corpus in its further pretraining. Just like its predecessor, Llama-2-Ko operates within the broad range of generative text models that stretch from 7 billion to 70 billion parameters. This repository focuses on the 70B pretrained version, which is tailored to fit the Hugging Face Transformers format. For access to the other models, feel free to consult the index provided below.

Model Details

Model Developers Junbum Lee (Beomi)

Variations Llama-2-Ko will come in a range of parameter sizes — 7B, 13B, and 70B — as well as pretrained and fine-tuned variations.

Input Models input text only.

Output Models generate text only.

Usage

Use with 8bit inference

Requires > 74GB vram (compatible with 4x RTX 3090/4090 or 1x A100/H100 80G or 2x RTX 6000 ada/A6000 48G)

from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline

model_8bit = AutoModelForCausalLM.from_pretrained(
    "beomi/llama-2-ko-70b", 
    load_in_8bit=True,
    device_map="auto",
)
tk = AutoTokenizer.from_pretrained('beomi/llama-2-ko-70b')
pipe = pipeline('text-generation', model=model_8bit, tokenizer=tk)

def gen(x):
    gended = pipe(f"### Title: {x}\n\n### Contents:",  # Since it this model is NOT finetuned with Instruction dataset, it is NOT optimal prompt.
        max_new_tokens=300,
        top_p=0.95,
        do_sample=True,
    )[0]['generated_text']
    print(len(gended))
    print(gended)

Use with bf16 inference

Requires > 150GB vram (compatible with 8x RTX 3090/4090 or 2x A100/H100 80G or 4x RTX 6000 ada/A6000 48G)

from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline

model = AutoModelForCausalLM.from_pretrained(
    "beomi/llama-2-ko-70b", 
    device_map="auto",
)
tk = AutoTokenizer.from_pretrained('beomi/llama-2-ko-70b')
pipe = pipeline('text-generation', model=model, tokenizer=tk)

def gen(x):
    gended = pipe(f"### Title: {x}\n\n### Contents:",  # Since it this model is NOT finetuned with Instruction dataset, it is NOT optimal prompt.
        max_new_tokens=300,
        top_p=0.95,
        do_sample=True,
    )[0]['generated_text']
    print(len(gended))
    print(gended)

Model Architecture

Llama-2-Ko is an auto-regressive language model that uses an optimized transformer architecture based on Llama-2.

	Training Data	Params	Content Length	GQA	Tokens	LR
Llama-2-Ko 70B	A new mix of Korean online data	70B	4k	✅	>20B	1e^-5
*Plan to train upto 300B tokens

Vocab Expansion

Model Name	Vocabulary Size	Description
Original Llama-2	32000	Sentencepiece BPE
Expanded Llama-2-Ko	46592	Sentencepiece BPE. Added Korean vocab and merges
*Note: Llama-2-Ko 70B uses `46592` not `46336`(7B), will update new 7B model soon.

Tokenizing "안녕하세요, 오늘은 날씨가 좋네요. ㅎㅎ"

Model	Tokens
Llama-2	`['▁', '안', '<0xEB>', '<0x85>', '<0x95>', '하', '세', '요', ',', '▁', '오', '<0xEB>', '<0x8A>', '<0x98>', '은', '▁', '<0xEB>', '<0x82>', '<0xA0>', '씨', '가', '▁', '<0xEC>', '<0xA2>', '<0x8B>', '<0xEB>', '<0x84>', '<0xA4>', '요', '.', '▁', '<0xE3>', '<0x85>', '<0x8E>', '<0xE3>', '<0x85>', '<0x8E>']`
Llama-2-Ko *70B	`['▁안녕', '하세요', ',', '▁오늘은', '▁날', '씨가', '▁좋네요', '.', '▁', 'ㅎ', 'ㅎ']`

Tokenizing "Llama 2: Open Foundation and Fine-Tuned Chat Models"

Model	Tokens
Llama-2	`['▁L', 'l', 'ama', '▁', '2', ':', '▁Open', '▁Foundation', '▁and', '▁Fine', '-', 'T', 'un', 'ed', '▁Ch', 'at', '▁Mod', 'els']`
Llama-2-Ko 70B	`['▁L', 'l', 'ama', '▁', '2', ':', '▁Open', '▁Foundation', '▁and', '▁Fine', '-', 'T', 'un', 'ed', '▁Ch', 'at', '▁Mod', 'els']`

Model Benchmark

LM Eval Harness - Korean (polyglot branch)

Used EleutherAI's lm-evaluation-harness https://github.com/EleutherAI/lm-evaluation-harness/tree/polyglot

TBD

Note for oobabooga/text-generation-webui

Remove ValueError at load_tokenizer function(line 109 or near), in modules/models.py.

diff --git a/modules/models.py b/modules/models.py
index 232d5fa..de5b7a0 100644
--- a/modules/models.py
+++ b/modules/models.py
@@ -106,7 +106,7 @@ def load_tokenizer(model_name, model):
                 trust_remote_code=shared.args.trust_remote_code,
                 use_fast=False
             )
-        except ValueError:
+        except:
             tokenizer = AutoTokenizer.from_pretrained(
                 path_to_model,
                 trust_remote_code=shared.args.trust_remote_code,

Since Llama-2-Ko uses FastTokenizer provided by HF tokenizers NOT sentencepiece package, it is required to use use_fast=True option when initialize tokenizer.

Apple Sillicon does not support BF16 computing, use CPU instead. (BF16 is supported when using NVIDIA GPU)

LICENSE

Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International Public License, under LLAMA 2 COMMUNITY LICENSE AGREEMENT
Full License available at: https://huggingface.co/beomi/llama-2-ko-70b/blob/main/LICENSE
For Commercial Usage, contact Author.

Citation

@misc {l._junbum_2023,
    author       = { {L. Junbum} },
    title        = { llama-2-ko-70b },
    year         = 2023,
    url          = { https://huggingface.co/beomi/llama-2-ko-70b },
    doi          = { 10.57967/hf/1130 },
    publisher    = { Hugging Face }
}

Acknowledgement

The training is supported by TPU Research Cloud program.

beomi
/

llama-2-ko-70b

Access Llama-2-Ko on Hugging Face